Generality of WM Capacity 1 The Generality of Working-Memory Capacity: A Latent-Variable Approach to Verbal and Visuo-Spatial Memory Span and Reasoning

A latent-variable study examined whether verbal and visuo-spatial working-memory (WM) capacity measures reflect a primarily domain-general construct by testing 236 participants in 3 span tests each of verbal WM, visuo-spatial WM, verbal short-term memory (STM), and visuo-spatial STM, as well as in tests of verbal and spatial reasoning and general fluid intelligence (Gf). Confirmatory factor analyses and structural-equation models indicated that the WM tasks largely reflected a domaingeneral factor that was a strong predictor of Gf and a weaker predictor of domain-specific reasoning. Variance from span tasks associated with verbal storage and rehearsal, rather than executive attention, predicted verbal reasoning, whereas variance associated with spatial storage and rehearsal predicted both spatial reasoning and Gf. The findings support a domain-general view of WM capacity in which executive-attention processes drive the broad predictive utility of WM span measures. Generality of WM Capacity 3 The generality versus specificity of cognitive abilities, mechanisms, and structures has triggered lively debate throughout psychology’s history, for example surrounding questions of general versus multiple intelligences (e.g., Guilford, 1967; Jensen, 1998; Spearman, 1927; Thurstone, 1938), single versus multiple pools of attentional resources (e.g. Kahneman, 1973; Navon, 1984; Navon & Gopher, 1979; Wickens, 1980, 1984), and process versus systems approaches to long-term memory (LTM; Toth & Hunt, 1999; Tulving, 1985, 1999; Weldon, 1999). Likewise, cognitive psychology’s two most prominent models of immediate memory differ most markedly in their proposals about the unitary versus fractionated nature of the construct. Atkinson and Shiffrin (1968) proposed a unitary short-term store (STS) that was specialized for holding information in a speech-based code, and although they allowed that further research might elucidate additional codes, all were thought to operate within the monolithic STS. In contrast, Baddeley (1986, 2000) proposes a partitioned “working memory” (WM) system with two domain-specific storage structures: a phonological loop that is specialized for maintaining verbal/linguistic information, much like the STS, and a visuo-spatial sketchpad that is specialized for maintaining visual and spatial information. Evidence for the distinction between verbal and visuo-spatial storage comes from numerous empirical dissociations in dual-task and neuropsychological/neuroimaging studies (for reviews see Baddeley & Logie, 1999; Henson, 2001; Jonides et al., 1996; Logie, 1995). However, some domain generality is proposed within Baddeley’s (1986; 2000) multicomponent model: A central executive acts as an attention-control structure and coordinator for the two storage components and their interaction with LTM. This domain-general executive provides the WM model with its functional emphasis as a system that maintains information in the service of ongoing cognitive activity such as language comprehension, imagery, and reasoning. That is, the dynamic interaction of domain-specific storage with the domain-general central executive gives the WM model the power to address the role of active memory in “real world” cognition. One might expect, then, that individual-differences research focusing on the relation between WM and aspects of complex cognitive ability would assume a mixture of domain-specific and domainGenerality of WM Capacity 4 general WM contributions. However, this has not always been the case. Daneman and Carpenter (1980, 1983) invented the first viable measure of WM capacity (WMC) and argued that it selectively engaged the central executive component of WM in a domain-specific fashion. Their “reading span” task presented short series of sentences for comprehension, followed by a memory test for all the sentence-final words in the series. Reading span is therefore a dual task, incorporating traditional memory-span demands with a secondary processing task that putatively engages the central executive. What created widespread interest in reading span (and related WM span tasks), was Daneman and Carpenter’s finding – now extensively replicated – that reading span strongly predicted comprehension abilities in ways that simple STM storage tasks did not (for reviews, see Daneman & Merikle, 1996; Carpenter, Miyake & Just, 1995; Just & Carpenter, 1992). Daneman, Carpenter, and their colleagues (e.g., Carpenter & Just, 1988; Daneman & Carpenter, 1980; Daneman & Tardif, 1987) originally hypothesized that the resource trade-off between processing and storage in the reading span task is specific to language comprehension. By this domain-specific view, reading-span performance is primarily a function of reading efficiency. With more efficient reading skills, good readers consequently have greater WMC remaining to store the products of that processing. Accordingly, span measures of WMC (i.e., the number of products effectively stored) are necessarily tied to a specific processing task, such as reading: The relatively large WMC of good readers is peculiar to reading comprehension tasks. In task domains where good readers are less skilled, their residual storage capacity during processing should be relatively limited. The strongest evidence for such domain specificity in WMC has come from studies contrasting the utility of span tasks consisting of verbal versus spatial materials for predicting complex verbal versus spatial abilities. If WM span tasks measure domain-specific capacities, then verbal span tasks should have limited value in predicting spatial ability, and vice versa. Indeed, Daneman and Tardif (1987) first reported that, whereas span tasks using verbal and numerical materials correlated significantly with verbal ability measures (rs = .61 and .51, respectively), a spatial span task did not (r = -.09). Using different tasks, Morrell and Park (1993) similarly found that verbal and numerical span Generality of WM Capacity 5 tasks, but not spatial span tasks, predicted standardized measures of text comprehension, and only spatial span predicted object assembly performance from diagrammatic, visuo-spatial instructions. Shah and Miyake (1996) reported the most influential and compelling demonstration of the verbal/spatial dissociation in WMC. In one study, reading span predicted verbal Scholastic Aptitude Test (SAT) scores more strongly than did a spatial task requiring mental rotation of letters and memory for their orientations (rs = .45 and .07, respectively). Conversely, spatial span predicted a composite of standardized visuo-spatial tests better than did reading span (rs = .66 and .12). In an exploratory factor analysis the spatial span/ability tests yielded one factor and reading span/verbal SAT yielded another. Similar findings with these same span tasks have been reported by other researchers, with reading span predicting only verbal performance and rotation span predicting only spatial performance (Friedman & Miyake, 2000; Handley, Capon, Copp & Harper, 2002). In a second study, Shah and Miyake (1996) crossed the domains of the processing and storage components to create two span tests each with matching and mismatching domains. For the matching tasks, one task had both verbal processing and storage stimuli and another had both spatial processing and storage stimuli (reading-word span, rotation-arrow span). The other two tasks had domain-mismatching processing and storage stimuli (reading-arrow span and rotation-arrow span). Here, domain of the storage items (words versus arrows), rather than the processing items, most strongly influenced the correlations with verbal and spatial ability measures. For example, spatialability scores were best predicted by rotation-arrow and reading-arrow tasks (rs = .68 and .65, respectively) compared to rotation-word and reading-word tasks (rs = .02 and .17). Although partialcorrelation analyses indicated that the processing domain of the tasks had a small effect on the correlations, domain-specificity in WMC tasks seemed to be driven primarily by the memory items, indicative of domain-specific storage mechanisms at work, leaving open the possibility that the executive contributions to these span tasks are more domain-general in nature. Despite findings that seem to indicate the domain-specificity of WMC, we have proposed a more domain-general view, which is also more consistent with the Baddeley model’s (1986; 2000) Generality of WM Capacity 6 proposal of a domain-general executive interacting with domain-specific storage structures/processes. Our view is that WM span tasks involve joint contributions of a general executive-attention capability and domain-specific rehearsal, coding, and storage processes. Of importance, the shared variance among measures of WM span and complex cognition reflects primarily the contribution of domaingeneral attention control, rather than domain-specific storage/rehearsal (e.g., Engle, 2001, 2002; Engle & Kane, in press; Engle, Kane & Tuholski, 1999; Kane, Bleckley, Conway & Engle, 1999; Kane & Engle, 2002). We argue that the critical, “executive-attention,” capability is one by which memory representations – for action plans, goal states, or environmental stimuli – are maintained in a highly active and easily accessible state. WM span tasks elicit such active maintenance by providing proactive interference that accumulates over trials presenting similar stimuli, thus making retrieval from inactive LTM difficult and slow (May, Hasher & Kane, 1999; Lustig, May & Hasher, 2001). Simultaneously, WM span tasks challenge that active maintenance by imposing shifts of attentional focus, to the unrelated secondary task, between the presentations of each to-be-remembered item. So, when we use the term “WMC,” we really mean the domain-general executive component of the WM system. Our view is that correlations between WM span and complex cognition are jointly determined by general executive-attention and domain-specific storage, but primarily by executive attention. Thus, a WMC measure should be quite general in predicting cognitive function. That is, the memory span test could be embedded in a secondary processing task that is unrelated to any particular skill or ability and still predict success in a higher-level task. Evidence supporting this view comes from three sources: 1) manipulating the processing demands of verbal WM span tasks and noting their relations to comprehension; 2) examining the link between verbal WM span and measures of general fluid intelligence; 3) examining the link between verbal WM span and low-level attention capabilities. To review these findings briefly, WM span tasks that require either reading comprehension or equation solution as the secondary task account for the same variance in measures of reading comprehension, indicating that reading per se does not drive the correlation between WM span and Generality of WM Capacity 7 comprehension (e.g., Turner & Engle, 1989). Moreover, skill in the secondary processing task has no apparent impact on the correlation between WM span and comprehension: Statistically accounting for participants’ strategic allocation of processing time to the secondary task, or matching the difficulty of the secondary task to each participant’s skill level, does not diminish the correlation between WM span and verbal-ability measures (Conway & Engle, 1996; Engle, Cantor & Carullo, 1992). Recent work using more sophisticated latent-variable methods has demonstrated further a strong relation between verbal WMC measures and non-verbal tests of general fluid intelligence (Gf), such as Ravens Progressive Matrices (Raven, Raven & Court, 1998). Here, the variance common to WM span tests representing different verbal/symbolic domains was statistically extracted to yield a pure measure, or “latent variable,” of the WMC construct. That is, measurement error associated with the task-specific variance in reading comprehension (reading span), mathematical ability (operation span) and enumeration (counting span) was partialed out, leaving only the variance shared among all the tasks. This latent variable correlated substantially (approximately .60) with the variance shared among prototypical Gf tasks, while a latent variable comprised of verbal STM tasks accounted for no unique variance in Gf (Conway, Cowan, Bunting, Therriault & Minkoff, 2002; Engle, Tuholski, Laughlin & Conway, 1999; see also Kail & Hall, 2001). Thus, it appears that WMC, as measured across tasks using a variety of symbolic materials, demonstrates predictive power well beyond the verbal domain and represents an important aspect of general intellectual ability. Quasi-experimental research has shown further that individuals identified as having high or low WMC, based on a verbal WM span task, differ significantly in their performance on attention tasks that bear little surface similarity to span tasks or each other. For example, in a dichotic-listening task, in which participants’ names were spoken once in the distractor ear, high-span participants were much less likely to report hearing their names than were low spans, indicating that high spans more effectively blocked the processing of the irrelevant message (Conway, Cowan & Bunting, 2001). In an antisaccade task, in which participants had to quickly look away from an attention-capturing cue in order to identify a pattern-masked letter appearing in an opposite location, high spans were less error Generality of WM Capacity 8 prone in blocking eye movements toward the cue, faster in making correct eye movements, and faster in correcting eye-movement errors, than were low spans (Kane, Bleckley, Conway & Engle, 2001). Finally, in a Stroop (1935) task in which participants had to name the color of conflicting color words, high spans were either faster or less error prone than were low spans, depending on task context (Kane & Engle, 2003; Long & Pratt, 2002). Verbal measures of WMC thus predict performance on simple attention tasks, presenting either verbal or non-verbal stimuli, either visually or aurally. Of importance is that all these attention tasks demand that participants block a pre-potent response in favor of a novel goal-directed one, in our view by using executive attention to actively maintain the goal (see also De Jong, 2000, 2001; Duncan, 1995; Roberts and Pennington, 1996; West, 2001; for a related but alternative view, see Hasher, Zacks & May, 1999). We see the diverse findings reviewed above as compelling evidence that WMC reflects primarily a domain-general, attentional construct that is important to a range of intellectual abilities. Our view thus conflicts with the findings discussed previously that suggest a strong dissociation between verbal and visuo-spatial WMC and reasoning (Daneman & Tardif, 1987; Friedman & Miyake, 2000; Handley et al., 2002; Morrell & Park, 1994; Shah & Miyake, 1996). However, we believe there are good reasons to be skeptical of the evidence for a strong domain-specificity in WMC. First, the “domain-specific” studies have limitations related to their participants. Several studies tested relatively small, and potentially quite homogeneous, samples (i.e., students at selective universities; Daneman & Tardif, 1987; Friedman & Miyake, 2000; Handley et al., 2002; Shah & Miyake, 1996). The concern here, from a psychometric perspective, is that when participants are sampled from a restricted range of general ability, general ability can only make limited contributions to any observed correlations. Instead, the variability that is detected, even if substantial, must be attributed to primarily domain-specific skills, strategies, and abilities (e.g., Deary, Egan, Gibson, Austin, Brand & Kellaghan, 1996; Detterman & Daniel, 1989; Legree, Pifer & Grafton, 1996; Spearman, 1927). Thus, domain generality may have been underestimated in these studies. As a case in point, Shah and Miyake (1996) tested mostly undergraduates from the highly selective Generality of WM Capacity 9 Carnegie-Mellon University, and reported low correlations between verbal and quantitative SAT scores in their samples (rs = .28, .45, and .59), relative to those seen in more comprehensive state universities with diverse student bodies (e.g., rs = .54, .69 and .74 in Turner & Engle, 1989; Engle et al., 1992; Engle, Tuholski et al., 1999, respectively). Such low correlations suggest a limited range of general ability within their participants and an artificially strong influence of domain-specific abilities. In fact, Shah and Miyake (1996) explicitly recognized this issue, noting that their results, “might not be generalizable to more cognitively diverse samples, such as those including non-college students, young children, or elderly adults. The dissociation between spatial and verbal measures, for example, could be less apparent among such samples, in which individuals are likely to vary more widely in general ability factors” (p. 21). As Shah and Miyake noted further, the benefits of studying ability-restricted samples is that the imposed limitation of domain-general contributions to performance can make it easier to measure and study the particulars of domain-specific contributions and mechanisms. We agree that if one’s goal is to examine the workings of domain-specific processes underlying WM span tasks, which do appear to exist, then studying a restricted range of general ability may well be a good strategy. However, if one’s goal is to assess the extent to which domain-specific and domain-general processes jointly contribute to WM span, then one must study a population that has ample variability in both general and domain-specific cognitive ability. A further obstacle to interpreting reported verbal/spatial WMC dissociations is that some of the verbal and spatial tasks differed dramatically in their difficulty. For example, the verbal and numerical tasks used by Daneman and Tardif (1987), in contrast to the spatial task, required much higher levels of domain-specific knowledge than is typical of WM span tasks (e.g., the processing component of the verbal task required generation of low-frequency words such as “sinewed”). These knowledge requirements can be inferred from the mean processing accuracies in the verbal (78%) and numerical (87%) tasks versus the spatial (96%) tasks, as well as from the fact that non-storage versions of these processing tasks correlated just as well with verbal ability as did their processing-plus-storage counterparts! It therefore seems likely that the domain-specific demands of these tasks may have Generality of WM Capacity 10 influenced their correlations with verbal ability, independent of their WMC demands. We find a similar problem with the tasks from Morrell and Park (1993), where the spatial WM task produced 50% lower span scores than the verbal WM tasks, obviously complicating the interpretation of cross-domain correlations. Although Conway and Engle (1996) demonstrated that processing-task difficulty within span tasks does not impact WM span/ability correlations, the difficulty range they explored was rather narrower than those seen here. A final difficulty in interpreting the dissociations between verbal and spatial WM span tasks is that some have been highly inconsistent. The reading span and rotation span tasks developed by Shah and Miyake (1996) to measure verbal and spatial WMC, respectively, demonstrate unstable correlations with each other across studies, with rs = .23 and .22 in Shah and Miyake (1996), and Handley et al. (2002), respectively, and rs = .42 and .04 in Friedman and Miyake (2000) experiments 1 and 2, respectively [slightly modified versions of these tasks by Sohn and Doane (2002) correlated at .42]. These varying correlations make it difficult to ascertain whether the tasks truly measure overlapping or non-overlapping constructs. A related, but broader, interpretive problem with this literature is that all the studies that have reported strong dissociations between verbal and spatial WMC have used a single task to measure each construct of interest. Because all cognitive tasks reflect multiple processes, we cannot know whether the observed dissociations in these studies reflect the domain-specificity of the WMC construct, or instead, the domain-specificity of non-WMC-related processes that also contributed to scores (i.e., measurement error). In addition to the ambiguities surrounding the domain-specific findings, we find four categories of data that directly suggest the generality of WMC across verbal and visuo-spatial domains. First, verbal WM span can sometimes predict spatial ability, and spatial WM span can sometimes predict verbal ability, with cross-domain correlations as high as those within domains (Bayliss, Jarrold, Gunn & Baddeley, 2003; Salthouse, Babcock & Shaw, 1991; Salthouse & Mitchell, 1989; Salthouse, Mitchell, Skovronek & Babcock, 1989; Swanson, 1996; Swanson & Howell, 2001; Süß, Oberauer, Wittmann, Wilhelm & Schulze, 2002). Second, cross-domain correlations among WM span tasks are Generality of WM Capacity 11 higher than those among STM span tasks, suggesting that STM reflects more domain-specific skills, strategies, and storage abilities than does WMC (Babcock & Salthouse, 1990; Henry, 2001; Park, Lautenschlager, Hedden, Davidson, Smith & Smith, 2002; Swanson & Howell, 2001). Third, individual differences in domain-specific ability can be substantially reduced by accounting for WM span measured in a different domain (Salthouse et al., 1989; Swanson & Sachse-Lee, 2001; Wilson & Swanson, 2001). Fourth, studies using latent-variable procedures find that constructs comprised of multiple verbal and spatial WM span tasks share ≥ 65% of their variance (Ackerman, Beier & Perdue, 2002; Kyllonen, 1993; Law, Morrin & Pellegrino, 1995; Oberauer, Süß, Schulze, Wilhelm & Wittmann, 2000; Oberauer, Süß, Wilhelm & Wittmann, 2003; Park et al., 2002; Salthouse, 1995; Süß et al., 2002; Swanson, 1996; Wilson & Swanson, 2001). In the present study we take a latent-variable approach to the question of WMC generality and its relation to verbal and visuo-spatial reasoning abilities, and we do so because it allows more definitive conclusions about underlying constructs than do simple correlational approaches. Latentvariable procedures require multiple tasks to measure each construct, and they statistically remove the error variance associated with the individual, imperfect tasks, retaining only the variance shared among all the tasks. This shared variance represents the latent construct of interest. Construct measurement based on multiple indicators is therefore more reliable and valid than that based on a single, multiply determined task. When using only a single task per construct, one cannot know which processes engaged by the task are responsible for observed correlations (the same problem exists with respect to interpreting experimental dissociations; see Jacoby, 1991; Toth & Hunt, 1999). In this regard, then, latent-variable techniques are analogous to the converging-operations approach to experimental research (see Salthouse, 2001), where concepts are operationally defined by using several imperfectly correlated conditions to eliminate alternative hypotheses (Bridgman, 1945; Garner, 1954; Garner, Hake & Eriksen, 1956). For example, constructs such as the phonological loop are validated experimentally by manipulating several task variables, such as articulatory suppression, irrelevant background speech, and word length (see Baddeley, 1986). In a similar manner, Generality of WM Capacity 12 psychometric constructs such as WMC are best measured using multiple tasks that differ in some surface respects, but which share theoretically critical requirements or processes. Moreover, latentvariable techniques, unlike correlational, regression, or exploratory factor analyses, permit hypothesis testing by statistically contrasting the fit of a priori theoretical models (factor structures) to the data. Here, then, we used multiple measures of verbal and visuo-spatial WM span, and statistically tested whether a single WMC factor, representing a domain-general construct, fit the data as well as two domain-specific WMC factors (and if the latter provided a better fit, how strongly the verbal and spatial factors were correlated). In contrast to most prior studies of verbal versus visuo-spatial WMC, we also included multiple measures of verbal and visuo-spatial STM, so that the generality of WMC could be compared to that of STM and the contributions of STM could be partialed out of the correlation between WMC and complex reasoning in the verbal and spatial domains. Including measures of verbal and spatial STM also allowed us to investigate the recent finding that STM and WMC may be less separable in the spatial domain than in the verbal domain, that is, that spatial tests of STM may tap executive-control processes in ways that verbal STM tasks do not (Miyake, Friedman, Rettinger, Shah & Hegarty, 2001; Oberauer, 1993; Oberauer et al., 2000; Shah & Miyake, 1996). Such findings indicate that span tasks need not have a dual-task requirement to reflect the WMC construct, which certainly complicates the field’s efforts to determine what the WMC construct actually represents (see Oberauer, in press). Also in contrast to prior work, we used verbal and visuo-spatial span tasks that shared a common procedure. In several verbal/spatial WMC studies (Oberauer et al., 2000; Oberauer et al., 2003; Swanson, 1996; Süß et al., 2002; Wilson & Swanson, 2001), WM tasks varied from the traditional complex-span procedure, to Brown-Peterson-like tasks in which groups of memory items are followed by a single rehearsal-prevention task, to “coordination” tasks that require the values of memory items to be updated on-line per particular transformation instructions, to immediate recall tasks presenting a large amount of material to study and remember (such as a paragraph or street map). The benefit of such broad task-selection strategies is that the latent variable derived from these Generality of WM Capacity 13 tasks will not include method-specific variance. However, the attendant costs are: (1) imprecision in understanding what the latent variable represents; that is, the variance shared by WM span, BrownPeterson, coordination, supra-span-recall, and other memory tasks may not reflect the same construct that has been traditionally defined by WM span tasks; (2) difficulty in insuring that task/method differences are appropriately balanced between verbal and spatial domains; (3) allowing that the WM/Gf correlation merely reflects that one broadly defined general factor can predict another; that is, a strong correlation between a Gf factor and a more circumscribed WMC factor is more impressive (i.e., more compellingly reductionistic) than is a similar correlation between Gf and WMC factors that each reflect a similarly diverse array of task types. Here we chose clarity of interpretation over breadth of measurement, and modeled all the verbal and spatial WM tasks after reading span, where the presentation of memory items is interleaved with a processing task and the lengths of the item lists vary across trials. Performance in these tasks, with either verbal or visuo-spatial materials, reflects one’s ability to encode, maintain, and retrieve lists of isolated stimuli in the face of a regularly occurring, highly interfering distractor task. Thus, if we were to find a close relation between WMC and Gf using these circumscribed tasks, it would not be because we simply substituted one poorly understood general factor for another. Yet an additional benefit of the span procedure is that we could create STM versions of each of our WM tasks that presented the same to-be-remembered stimuli, but without the additional processing demand of the secondary task. Here we also tested a relatively broad sample of participants, in order to best allow domaingeneral contributions to WMC to be measured. We did so by recruiting participants from a variety of university populations (i.e., from a highly selective technical institute and two comprehensive state universities) and also a subset from two community populations. Although most of our community volunteers had at least some college background, some did not, and a small proportion had not graduated from high school. Thus, with a modestly diverse sample with respect to education, culture, Generality of WM Capacity 14 and race, we hoped to better gauge the domain-generality of WMC than was possible by much of the previous research. At one level, it might be unsurprising for us to find that verbal and visuo-spatial WMC tasks were highly correlated. After all, if the tasks are identical aside from the particular stimuli used, high correlations based on method variance, alone, might be expected. However, a clear demonstration of domain-generality in WMC here would be important for a number of reasons. First, as we reviewed above, there are several published reports of strong verbal/spatial WMC dissociations that are influencing WM theory, we believe inappropriately. Recent reports of WMC generality across verbal and spatial domains have appeared in specialized educational, developmental, or psychometric journals, and so they are unfamiliar to many experimental psychologists. Second, and somewhat counterintuitively, we predict that verbal and spatial WM span measures will be more strongly intercorrelated, due to their general executive contributions, than will verbal and spatial STM measures, despite the fact that the STM measures are more methodologically similar to one another than are the WMC measures. Thus, a finding of substantial domain generality of WMC, but not of STM, will powerfully indicate generality at the underlying construct level rather than the surface, measurement level. Third, and of most importance, WM is now a central construct in cognitive theory, and so the WMC literature impacts basic cognitive, intelligence, developmental, and neuroscience research. WMC also is increasingly influencing applied domains, particularly regarding interpersonal, clinical, and health-related issues (e.g., Arnett et al., 1999; Brewin and Beaton, 2001; Clark et al., 2003; Finn, 2002; Richeson & Shelton, 2003; Rosen et al., 2002). Understanding what, exactly, WM span tasks measure, and what the WMC construct ultimately reflects, is therefore critical to several lines of psychological inquiry. If WMC, as widely measured by verbal span tasks, is a construct relevant only to verbal/symbolic domains, this should radically affect the way theorists interpret its close relations to attention, self-control, intelligence, neuropsychological diagnoses, and executive function. Method